Key Takeaways from the Morning

  • We turned Bronze data (emails) into Silver data (structured records)
  • Created a queryable record of all customer interactions
  • Added AI to draft intelligent responses

Next: But what about existing unstructured data? That's next...

Key Takeaways from the Morning
Slide 1

What About Unstructured Data?

We've built a system to capture structured data going forward. We've also spent time looking at ways of structuring our data – turning bronze into silver – making it ready to work with AI.

But what about existing unstructured data? Can our AI tools help extract useful, strategic insights from:

  • Old emails
  • PDFs in folders
  • Word documents
  • Web research, etc.

Good news: AI tools can help extract structure from the mess. They can glean meaning from tangled documents.

Slide 2
Part 1 – An Example Workflow

Part 1 – An Example Workflow

Previously, when working with numerical data, 100% accuracy was required to generate meaningful insights, especially for financial reporting or forecasting. That data needed to be analytics-ready: correct, stable, and suitable for dashboards and reports.

This afternoon's data is text-based and more diverse in form and purpose. It is not worthless – there is gold hidden among the bronze. Important decisions, lost details in email chains, and useful trends often live here.

Manually reviewing large volumes of text would be extremely time-consuming.

This is where AI can step in.

Slide 3

The Data Sources

This example demonstrates an advanced workflow used to compare an AI bootcamp curriculum with official UK government guidance on AI workforce skills.

The data sources included:

  • Bootcamp materials: Word documents, PDFs, transcripts, slides, schemes of work, blog posts, reflection notes
  • Government guidance: 7 web pages from the gov.uk AI Skills for the UK Workforce report (October 2025)

The goal: Identify alignment, find gaps, and create an actionable improvement roadmap.

Slide 4

Step 1 – Data Standardisation

Quick technical note: All sources were converted into a single data format: Markdown (.md).

Why this matters:

  • Different file types have different structures
  • Plain text with markdown formatting creates uniformity
  • AI tools can process all sources consistently
  • Web pages were scraped and cleaned; documents were converted using standard tools

This is the bronze-to-silver data cleaning principle in action.

Step 1 – Data Standardisation
Slide 5

Step 2 – Chunked Analysis (The Core Technique)

Rather than analysing everything at once, a chunked approach was used:

  1. AI reads all bootcamp materials → creates a synthesis document
  2. AI processes each government page separately
  3. AI compares one government page at a time against the synthesis
  4. AI synthesises all analyses into final reports

This approach reduces failure points and allows human check-ins.

Why chunking works: Each source gets proper attention, patterns emerge naturally, context limits are avoided, and understanding builds incrementally.

Slide 6
Step 3 – Structured Outputs

Step 3 – Structured Outputs

Each analysis followed a consistent structure:

Key Government Recommendations

Alignment with Current Approach
- Strong alignments
- Partial alignments

Gaps Identified

Quick Wins

This made results quick to scan and easy to compare across sources.

Slide 7

What I Learned

The analysis showed strong alignment with government guidance, while also revealing opportunities for improvement and new material. Unexpected trends emerged, including common learner questions and recurring technical challenges.

This work would normally require:

  • Enterprise analysis software
  • Data science expertise
  • Weeks of manual reading and comparison

Instead, it took a couple of hours with an AI coding assistant, structured prompts, and organised thinking—producing 16 analysis documents and a prioritised action plan.

Once the prompts were built, the AI worked autonomously with only periodic check-ins.

Slide 8

A Simpler Example

Same principles, simpler tools.

You do not need software engineering tools to do this. We will look at unstructured Word documents—summary emails sent to learners after past bootcamp sessions—and use browser-based AI tools to extract insight.

Live demo: ChatGPT, Gemini, Copilot, and NotebookLM.

A Simpler Example
Slide 9

NotebookLM Data Tables

NotebookLM’s Data Tables feature is purpose-built for turning unstructured data into reliable, queryable tables.

  • Upload unstructured sources (PDFs, documents, websites)
  • AI extracts structured data automatically
  • Queryable tables are created without manual data entry

Using NotebookLM’s research tools, the platform can also gather data for you.

We will now carry out a short competitor research task and explore the material NotebookLM gathers.

Slide 10